Fifteen forecasts for the end of 2026
Fifteen yes-or-no questions on AI, money, health, energy and policy, all settling by 31 December. Each has the odds I gave it, the public source that decides it and, where one existed, the prediction market's price. Scores go here as each one settles.
- What
- Fifteen questions that settle between 30 November and 31 December 2026, each with a named public source and a default if that source stays silent.
- Odds
- First set on 6 October 2026 in a private record, then updated on 5 November for the month's news. The 5 November odds are the ones scored; the 6 October odds are shown beside them.
- Scoring
- Brier score: the squared gap between the odds and what happened, averaged. Lower is better. Always saying 50% scores 0.250. The six questions with a prediction market are also scored against the market's price on 5 November.
- So far
- None has settled yet. The first settles on 30 November; results are added here as each one does.
1The forecasts
Sorted by the date each one settles. Open a row for the exact wording and the source that settles it.
DeepSeek releases a newer Pro model30 November 13%
Does DeepSeek publicly release a Pro-tier model newer than DeepSeek-V4-Pro-0813 (for example V4.1-Pro or V5-Pro) by 30 November 2026? Flash, vision, OCR models and re-uploads don't count.
Odds: 50% on 6 October, 13% on 5 November. No Pro release in October (the newest model is a Flash, 17 October); Polymarket 13.5%.
Market on 5 November: Polymarket 13.5% (the same question).
Settled by: Polymarket's market on the same question, checked against DeepSeek's Hugging Face releases. Source
SF-02 · AI & compute
Adobe: AI shopping traffic at least doubles2 December 55%
Does Adobe's Cyber Monday 2026 recap report traffic from AI tools to US retail sites up at least 100% on a year earlier (on Cyber Monday itself, or season to date if no Cyber Monday figure is given)?
Odds: 55% on 6 October and 5 November. No new information; unchanged.
Market: none asks this question.
Settled by: Adobe's Cyber Monday 2026 press release. No if Adobe gives neither figure. Source
SF-13 · Work & commerce
US unemployment at 4.3% or higher in November4 December 40%
Is the US unemployment rate (U-3) for November 2026 at least 4.3%, as first published?
Odds: 33% on 6 October, 40% on 5 November. Kalshi’s “above 4.2%” price rose from 31¢ to 38¢ bid, 42¢ ask.
Market on 5 November: Kalshi 40% (“above 4.2%”, middle of bid and ask).
Settled by: BLS series LNS14000000, November value as first published. Not scored if unpublished by 31 December. Source
SF-12 · Work & commerce
Bank of Canada raises its rate on 9 December9 December 38%
Does the Bank of Canada raise its policy rate at its 9 December 2026 announcement?
Odds: 53% on 6 October, 38% on 5 November. The Bank raised its rate to 2.50% on 28 October; the market now leans to a December hold.
Market on 5 November: Polymarket 38% (a 25-point hike in December).
Settled by: The Bank of Canada's rate announcement. Any increase is a Yes. Source
SF-14 · Canada
US hemp-THC limits take effect on schedule12 December 65%
On 12 December 2026, is the federal redefinition of hemp (a total-THC standard and a 0.4 mg-per-container cap) in effect, with no further delay, repeal or replacement enacted by 11 December?
Odds: 65% on 6 October and 5 November. No second delay has been reported; unchanged.
Market on 5 November: Manifold 67% (hemp-THC products banned by the end of 2026; close, not identical).
Settled by: Federal laws enacted 3 September to 11 December 2026 (congress.gov). Yes if no such law exists. Source
SF-10 · Society & policy
Glama lists 135,000 AI-tool servers31 December 75%
Does the Glama registry of MCP servers (the plug-ins AI agents use to call outside tools) show at least 135,000 servers on 31 December 2026?
Odds: 50% on 6 October, 75% on 5 November. The count reached 116,321, rising about 600 a day since mid-October, against about 330 a day needed.
Market: none asks this question.
Settled by: The “N in the Glama Registry” count, from the archived copy nearest 31 December. Other directories don't count. Source
SF-01 · AI & compute
Three AI assistants in Apple's top 10 apps of 202631 December 45%
Do at least three general-purpose AI-assistant apps (ChatGPT, Gemini, Meta AI (Muse), Claude, Grok, Copilot, Perplexity, DeepSeek; not the Google app) appear in the top 10 of Apple's US list of the most downloaded free iPhone apps of 2026?
Odds: 45% on 6 October and 5 November. No new information; unchanged.
Market: none asks this question.
Settled by: Apple Newsroom's annual list (about 10 December last year). Not scored if Apple doesn't publish it by 31 December. Source
SF-03 · AI search
Stablecoins pass $320 billion31 December 5%
Is total US-dollar stablecoin supply at least $320 billion on DeFiLlama for 31 December 2026?
Odds: 35% on 6 October, 5% on 5 November. Supply fell to $294.5 billion and needs to rise 8.6% in eight weeks.
Market: none asks this question.
Settled by: DeFiLlama's stablecoin chart, the data point for 31 December. Source
SF-04 · Money & markets
Tokenized Treasuries back above $16 billion31 December 10%
Is rwa.xyz's tokenized US Treasury “Distributed Value” at least $16.0 billion at the reading closest to 31 December 2026?
Odds: 30% on 6 October, 10% on 5 November. Distributed Value is $14.28 billion and needs to rise 12%.
Market: none asks this question.
Settled by: rwa.xyz, Distributed Value (not Represented Value), archived reading nearest 31 December. Source
SF-05 · Money & markets
Pew: 70% of US teens use AI chatbots31 December 30%
Does Pew Research Center publish a new survey of US teens by 31 December 2026 reporting that at least 70% use or have used AI chatbots? (Its December 2025 figure was 64%.)
Odds: 45% on 6 October, 30% on 5 November. Pew’s 2026 teen reports so far repeat its late-2025 survey (64%); a new survey before year-end is uncertain.
Market: none asks this question.
Settled by: Pew Research Center only. No if Pew publishes no such figure by 31 December. Source
SF-06 · Culture & attention
FDA approves psilocybin therapy (COMP360)31 December 8%
Does the FDA approve COMP360, Compass Pathways' psilocybin treatment, for treatment-resistant depression by 31 December 2026?
Odds: 16% on 6 October, 8% on 5 November. Compass now expects to finish its filing by mid-November (3 November). Even the 1–2-month review a priority voucher allows would barely reach 31 December. Manifold’s question is broader: any psilocybin approval in 2026.
Market on 5 November: Manifold 16% (any psilocybin approval in 2026).
Settled by: FDA press release or Drugs@FDA, or a Compass filing. Source
SF-07 · Bio & health
FDA moves BPC-157 to its allowed list31 December 15%
Does the FDA place the peptide BPC-157 in Category 1 of its 503A compounding list (or announce that it will) by 31 December 2026? Statements without FDA action don't count.
Odds: 45% on 6 October, 15% on 5 November. Still no FDA action on BPC-157; Polymarket fell to 15%.
Market on 5 November: Polymarket 15% (the same question).
Settled by: Polymarket's market on the same question, checked against the FDA's 503A list. Source
SF-08 · Bio & health
Home Assistant reaches 785,000 active installs31 December 25%
Does Home Assistant's public analytics show at least 785,000 active installations on any reading dated on or before 31 December 2026?
Odds: 50% on 6 October, 25% on 5 November. 720,965 on 5 November; the recent pace, about 26,000 a month, reaches roughly 770,000.
Market: none asks this question.
Settled by: analytics.home-assistant.io (opt-in installs; the history is kept). Source
SF-09 · Energy & the physical world
Brazil's lower house passes its digital-markets bill31 December 10%
Does Brazil's Chamber of Deputies approve bill PL 4675/2025 (digital markets, under the competition authority CADE) in a plenary vote on the bill itself by 31 December 2026? Committee votes and votes on urgency requests don't count.
Odds: 15% on 6 October, 10% on 5 November. No movement since August, and Congress is in its post-election session.
Market: none asks this question.
Settled by: The Chamber's record for the bill. Source
SF-11 · Society & policy
Reddit and Google renew their AI data deal31 December 20%
By 31 December 2026, do Reddit or Google publicly announce a renewed, extended or replacement AI data-licensing deal between them? Reports citing unnamed sources don't count.
Odds: 25% on 6 October, 20% on 5 November. No announcement; the current deal reportedly runs into 2027.
Market: none asks this question.
Settled by: Reddit investor releases, earnings remarks or SEC filings, or a Google post. Source
SF-15 · Where AI gets its sources
2How the odds were set
These questions come from a forecasting system I built and run with Claude, Anthropic's AI model. Each week it looks for things that are small or dismissed today but could be large in one to five years, and keeps about sixty long-range questions on them, each with odds. Those take one to three years to settle, which is too slow to learn whether the odds mean anything. So each quarter the system also writes fifteen questions that settle within about ninety days, spread across its subjects. These are the first fifteen.
Every question names the public source that settles it, what counts and what doesn't, and the answer if the source says nothing. Before any odds were set, each candidate was checked against that source to make sure it hadn't already happened. Six candidates had already happened and a seventh looked near-certain, so all seven were dropped; four more were dropped as too unlikely to teach anything.
Where a prediction market (Polymarket, Kalshi or Manifold) asked the same thing, its price was the starting point, and any gap from it needed a stated reason. The peptide question sits five points under Polymarket's 50%, for example, because the FDA had not acted in the ten weeks after its advisory committee voted. The odds are whole percentages. I approved the set on 6 October.
On 5 November, before publishing, the same system updated each question for the month's news, with a one-line reason for every change; open a row to see it. Twelve of the fifteen moved. Two of the largest moves follow markets: DeepSeek released no new Pro model in October and Polymarket fell from 51.5% to 13.5%, and the FDA still hasn't acted on BPC-157, where Polymarket fell from 50% to 15%. The stablecoin drop follows the data: supply fell instead of growing.
3Why the publication date matters
The 6 October odds sit in a private, dated record, so nothing on this page proves when they were set. That is why the score uses the 5 November odds, whose first public record is this page and its archived copy.2 The 6 October column stays visible so that anyone can see how far a month of news moved them.
From here the questions and the 5 November odds are fixed. Each one is scored as written, including the ones that later look wrong.
4How they will be scored
Each question gets a Brier score, the squared difference between the odds and the outcome (1 for Yes, 0 for No).1 Saying 80% on something that happens scores 0.04; saying 80% on something that doesn't scores 0.64. The average across the fifteen is the headline number. Two benchmarks come with it: 0.250, what you get by saying 50% to everything, and, on the 6 questions that have one, the score of the market's price on 5 November. That second comparison is weak this quarter: on four of the six, the 5 November odds simply match the market, and two of the markets ask a close rather than identical question. It means more where the odds and the market disagree.
Fifteen questions can't separate skill from luck on their own. The record that matters is the series: a new set each quarter, each published before it settles, and scored as written. A question whose source disappears or never publishes is marked “not scored” and left out of the average, as each one's rules state.
These are forecasts about public events, not financial advice.
References
- [1]G. W. Brier. Verification of forecasts expressed in terms of probability. Monthly Weather Review 78(1): 1–3, 1950.
- [2]Internet Archive. Wayback Machine copies of this page, from 5 November 2026.
Citing this note
Revisions. First version, 5 November 2026. Results are added as each question settles. The questions and the 5 November odds never change.